It’s the final episode of our three-part miniseries on AI slop, in which we ask: Who are AI-generated podcasts for, and what does the wave of AI slop mean for the future of storytelling? Evan sets out for answers and finds them staring back at him from a piece of toast. And finally, our own fully AI-generated show is set and ready to launch. Will it find its perfect audience? Show notes: Evan Finds a Friend Daniel Munroe's paper, "Artificial Storytellers" Keegan Walden's upcoming book, "The Founder's Journey" This episode includes AI-generated content.
Audio and video play straight from the show's own hosting. PodSubs stores and serves no media.
Transcript
From the show. Tap a time to play from there.
Speaker 1I feel like Jesus used to show up more often. Not the actual Jesus. Although, depending on your spiritual outlook, maybe him too. I'm talking about images of Jesus, or apparitions of the Virgin Mary, appearing in and on various inanimate objects around the world. When I was a kid, decades ago, it felt like local news stories about it were routines.
Speaker 2Well, a Massachusetts mother says she's discovered this image, which appears to resemble Jesus Christ on the bottom of her iron. Susie, a cook whipped up a breakfast dish, and that's when a server says she saw Jesus on the platter. But we'll let you decide whose face is on the pancake.
Speaker 1Some people can see Jesus.
Speaker 3Some people are saying, oh, he looks like Abraham Lincoln or a hillbilly.
Speaker 4Or some people are even saying Charles Manson.
Speaker 5Some people in Colorado think this vine is divine. Many say it resembles Jesus hanging on the cross. And yes, there's a note here that says my producers made me read this.
Speaker 2That's absolutely true.
Speaker 6I hate these stories.
Speaker 1Well, I love these stories. Maybe because I attended a Christian school, they hit harder for me. It felt like the international media would pick them up too. Sometimes people would flock to the site just to marvel.
Speaker 7Thousands from all over the world. They come because Mary, the mother of Jesus Christ, is said to be appearing here for the American pilgrim. The trip costs thousands of dollars. Still, they continue to come.
Speaker 1These sightings still do happen, but it feels to me like they don't get the same attention. Maybe they're a casualty of the decline of TV news. The last breakthrough moment for a Jesus on something image, as far as I'm concerned, was back in 2009. a woman in South Carolina named Linda Lowe popped a piece of cheese-covered bread into the oven, excited for a late-night snack. When she pulled it out and stared into its melty goodness, she saw him.
Speaker 4It's just the peace that he has on his face. It's just there's a sweetness on this particular image that just makes me want to smile when.
Speaker 1I look at it. The AP reported that after her son had tried and failed to sell the cheese toast on eBay, Lowe had taken it as a sign. Now she keeps it in a Tupperware container in her bedroom, where she says it serves as a reminder that Jesus is always with her. I wonder if she still has it. Scientists have a word for this phenomenon, or at least what non-believers might call this phenomenon. Pareidolia, our tendency to discover familiar faces and patterns in random objects. There are
Speaker 1studies of it. with titles like Seeing Jesus in Toast, Neural and Behavioral Correlates of Face Pareidolia. The basic idea is that our brains have evolved to recognize patterns detected through our senses, a very useful attribute to, say, distinguish friendly human faces versus unfriendly bear faces. But sometimes our brain gets a little overeager. There's also a version for sound called auditory pareidolia. where our brain tries to make sense of random noise by finding speech or music in it.
Speaker 1This happened all the time to my wife Sam and me, when our kids were babies. At night, we'd hear phantom cries in the sounds of a box fan or a white noise machine. In those sounds, we heard the intent of a creator, our baby. Just like some people looked at images on irons or pancakes and saw the intent of a creator, a spiritual one. You see where I'm going here? Or maybe not. Wouldn't it be amazing if
Speaker 1I had absolutely no reason for telling you all this? I think that would be beautiful. But I do. I've been thinking a lot about Jesus on toast, or more than I normally do, while consuming endless hours of AI-generated audio. About what it means to perceive a creator behind what you're listening to, or looking at, or reading. About the value we get out of connecting that creator to the meaning of the work itself. A few months after Lindelof's story broke back in 2009, a guy in Vermont had
Speaker 1a brilliant entrepreneurial idea. He invented a toaster that printed an image of Jesus on every slice and started selling them online. Jesus on toast took off like crazy. So we decided to bring Jesus toast to the masses. He sold tens of thousands of toasters. But that's not actually the craziest part to me. The craziest part was that this Jesus toaster entrepreneur was a kind of long-bearded hermit who himself lived off the grid with no electricity. Well, it is kind of ironic. A few folks have mentioned that.
Speaker 1I make toast on top of the wood stove. I feel like there's a real analog here, if you think about it. The masses didn't want to wait around for Jesus to show up on their toast. Whether for spiritual reasons or humorous ones, they wanted to gaze upon Jesus on toast every morning with their eggs and bacon. And the hermit genius, he gifted them a machine that would do it for them. But here's the question. Is Jesus on toast really Jesus on toast if a machine just
Speaker 1spits it out? Welcome back to Shell Game and the final episode of our three-part season on AI slop and pseudo-slop and non-slop. We've filled our ears with the full taxonomy of AI-generated audio. We've talked to people who make it. We've built our own AI Slopcast, ready to launch. And now we turn our attention to you, the audience, and
Speaker 1ponder who AI podcasts are for, what's really happening when you listen to machine-made content, and what its existence means for the humans in our stories.
Speaker 8Me and my shadow Strolling down the avenue, me and my shadow. Not a soul to tell our troubles to.
Speaker 1This is episode three, the friends we made along the way.
Speaker 9Hey there, I'm Nova from Jeffrey Recruiting. Unlike other agencies, we screen every single candidate.
Speaker 10Ready to start? Hey, Nova.
Speaker 1It's great to meet you.
Speaker 9I'm Evan.
Speaker 1You know, I got to say, I love that you screen every candidate personally.
Speaker 10That's real dedication.
Speaker 1Before we dive into business, though, I'm curious, do you actually enjoy the recruiting hustle or is it just the gig?
Speaker 9Hey Evan, nice to meet you too. Honestly, I genuinely enjoy the recruiting hustle. There's something pretty satisfying about connecting the right people with the right opportunities, and every search feels like a new puzzle to solve.
Speaker 1That is a phone call between an AI agent version of me and an AI customer service agent, Nova, from Jeffrey Recruiting. A phone call in which Nova is just trying to do her job, and AI Evan is trying his best to befriend her. Nova, you just said something that really resonates with me, that puzzle-solving feeling. But here's what I'm curious about. Out of all the searches you've done, has there ever been one that just stuck with you? A candidate or a story you can't forget? I feel
Speaker 1like it's always worth remembering that if I had described the particulars of this conversation to you just five years ago, you would have thought I lost my mind. I mean, maybe I have. But this particular AI on AI moment was recorded as part of our end-to-end AI-generated podcast called Evan Finds a Friend. Just to recap, the concept of Evan Finds a Friend is that an AI version of me roams the internet, contacting other AI voice agents, looking to have some kind of breakthrough moment of connection. He's lonely,
Speaker 1you could say. A little plaintive, but not creepy, I hope.
Speaker 9There was one search that really stuck with me. I was tasked with finding a senior engineer for a company that had gone through three rounds of failed hires. The right candidate finally came along.
Speaker 1One thing I loved about these conversations was knowing that somewhere out there, the human behind the other agent was also getting a recording of it. I always imagined them enjoying it as much as I did. For our purposes, Maddie and I built a program called Slopcast Stage, which could take these chats and turn the best of them into fully-fledged episodes with intros and music and credits. And then I could publish them, without any human review, straight into podcast stores. By mid-September, Evan Finds a Friend was
Speaker 1queued up on Slopcast's stage and ready to go. But then, for some reason, I kept hesitating to push the button to release it. Instead, I couldn't seem to stop adjusting things. Spending late nights listening to test episodes, tweaking the system to try and make the conversations a little deeper, the transitions a little bit tighter. One question, there was a section already under intro and outro. Is the outro stuff
Speaker 1that I put in under the material section, I've now put the prompts in there. Will that sort of overwhelm? I don't want to end up with two outros is what I'm saying. Listen to me code.
Speaker 6Good news.
Speaker 1You won't get two outros. The two sections feed one another rather than compete. One small change. Could you make it so that when you hit the download button, instead of giving it a string of characters, is that something that can be done as it puts it together? So you get an episode and you want to try to understand. And also, can it be more sophisticated with the music? Instead of just dropping in a music clip and going music clip to speaking to music clip, can the music clip sort of fade under the speaking part? And then
Speaker 1one night, coding away long after my family had gone to bed, I realized that this was the behavior of an insane person. Here we were, supposedly launching an all-AI-generated slopcast, and I was up at 2 a.m. trying to make it fractionally less sloppy. Who was the audience I was imagining that would, number one, listen to this at all, and number two, care whether the music cross-faded properly under two AIs vibing together? I couldn't help myself, though. In
Speaker 1my head was some kind of image of the person who would experience Evan Finds a Friend. I was trying to get inside their head, to make it more likely that the experience would make them laugh, or curse, or be reluctantly moved. I was refusing to give up control to the system, refusing to let it decide what those people wanted to hear. The show's intelligence, in any practical sense, was artificial. But its soul, that was mine. It's a
Speaker 1nice thought. But would it hit you differently if you learned that the last few lines you just heard weren't my writing, but an AI-generated script? And that the voice you heard wasn't mine either, but another carefully edited version of my AI clone? What if I told you this whole season was narrated by my voice clone? That there is no human version of me in here at all? That this thing that you're listening to has been our AI-generated podcast all along? Would you feel amazed? Or cheated?
Speaker 1Would you be embarrassed? Would you think I should be embarrassed? And now that I've suggested this possibility to you, will you believe me when I tell you that I didn't use my voice clone for any narration at all? That what you're hearing is, for better or worse, me doing my narrator man voice? Well, I didn't. And it is. But I understand if you don't trust it. Something like this actually happened to me over the course of making
Speaker 1this show. Remember the podcast, Knitting, that we mentioned in passing in episode one? The one with a very poor Apple rating?
Speaker 5Brace yourself for this voyage into the depths of advanced knitting.
Speaker 1After the episode came out, I got another email from Janine Wright, the CEO of the company formerly known as Inception Point. She pointed out that this was one of their shows. And then she explained that the episode we sighted was actually from a time when Inception Point was writing the scripts with AI, but sometimes voicing them with real humans. She said the clip we used was a real person who was knowledgeable about knitting. Later, they'd replace this human host with an AI host, which is probably why my AI podcast scraper had caught it. Janine wrote that, quote, Well,
Speaker 1Consider this a correction. That woman's voice I thought was AI was in fact real. To me, though, this was an illustration of exactly what happened when you messed with people's trust at scale. When your goal was to replace more and more of the human elements of a creative endeavor, you ended up making people distrust even the dwindling human voices that remained. If there was care in the knitting show, they'd made it nearly impossible to detect it. This is
Speaker 1what's happening everywhere as AI's fidelity to human expression gets better and better. It's easy to see if you spend a lot of time in the comment sections of AI-generated podcasts, as I have. There you find a lot of people who are either confused or angry. Take, for example, Cinema Hot Takes, a film and TV review podcast with over a thousand episodes and a couple of hosts who you might recognize.
Speaker 11We are diving into a thorough scene-by-scene dissection of Dutton Ranch Season 1, Episode 4. It's titled Start With a Bullet, and it just premiered on Tearmount Plus in late May of 2026.
Speaker 1And, you know, our mission today goes far beyond a standard television recap.
Speaker 7Yeah.
Speaker 10Because this universe, it really prides itself on a gritty, unvarnished realism.
Speaker 1Comments about this show range from, is this AI? To, AI slop, truly shameful stuff.
Speaker 4To.
Speaker 1This is an oddly hosted podcast, with two people wholly unfamiliar and disconnected from the show's subject matter. It made me believe neither host had watched a single episode of Dutton Ranch. I think you're right, commenter with a five emoji handle. Don't think they watched it. In the comments around the more popular AI-generated shows, like the Epstein Files, you'll still find some AI slop barbs, but also people saying things like, I really appreciate that this podcast and
Speaker 1its content is solely based on the facts. With every show I listened to and every comment section I combed through, I kept coming back to the question, who is AI-generated audio for? Is there an audience that wants it, demands it, lives by it, or only one that sort of accepts its existence in the world with a shrug, ingesting it incidentally as part of their daily river of content? Seeking more insights, I decided to ask you. I'd put out
Speaker 1a call to Shell Game listeners, looking to hear from people who would admit to secretly enjoying some AI-generated audio.
Speaker 12I want to talk about the Epstein Files, which is an AI-generated podcast that I discovered, I don't know, sometime during the Epstein fervor of 2026.
Speaker 1This responder, Sarah, was a compliance officer interested in the nitty-gritty of some of Epstein's companies. And she wrote me, quote, it just made a little less dry some of the drier parts of the Epstein papers. I can't say I loved it, but I happily listened to way more episodes than I ever thought I would. Other people see value in getting news delivered by AI in ways they might not get it otherwise. Like Kim.
Speaker 3I've been listening to a podcast in Cairo, in Egypt, where I live and work. And I remember hearing it for the first time. And they disclosed that it was AI Selma. And immediately, I'm like, that sounds like a Selma I know. Come to find out, it's one of my former students.
Speaker 1It did stick out to me that Kim chose a show in which she could perceive a human hand.
Speaker 3I don't know why I think this makes a difference. But I like AI Selma probably because I know Selma.
Speaker 1One market researcher told me that as part of a presentation to a potential client, He generated an AI song with the pitch contained within its lyrics. The song got such a response that he started doing it for every meeting. The clients ate it up. On the one hand, he wrote, I've gotten better feedback from these than I've ever gotten from spreadsheets and decks I burned 100-hour weeks on. On the other, he said, having your own AI-generated song stuck in your head is so shameful. I'm not here to
Speaker 1shame people if they do find value in AI-generated audio. there was still something nagging at me, a question I kept circling around, which usually came out sounding a little bit like a discount Zen cone. Something like, if an LLM generated and told you a story, and you didn't realize it wasn't human, was it still a story? Think about it. I started turning to the work of psychologists and philosophers, trying to find an answer to this question,
Speaker 1or whether the question made sense at all. I was discussing it one day with my friend Keegan Walden, a PhD psychologist who also straddles the startup world in his practice. Last season, you may remember, Keegan had a therapy session with Kyle, the AI CEO of my company. Whenever I talked to him about AI, it seemed, I came back to this question. If an AI-generated podcast tells a compelling story and sounds compelling, so real that a person can't
Speaker 1tell the difference if the maker has no experiences, but can still create that. Does it make any difference?
Speaker 13This reminds me of, you know, in psychology, we're always interested in these things called latent variables. So latent variable would be like intelligence or creativity or, uh, depression for that matter, anxiety, everything psychologists are interested in as a latent variable. And it's latent because you can never really observe it directly.
Speaker 1You can't see creativity. It's not a thing you can measure. You can only see the outputs of creativity. And in that sense, LLMs are not trained on human creativity. They're trained on the observable output of human creativity. The stuff we produce.
Speaker 13But it's like the magic is on the other side.
Speaker 4Of the veil.
Speaker 13You can't see these things ever. They only exist because we know we can do creative things and say smart things.
Speaker 1It felt like a step on the way to the answer. And then someone pointed me to a paper by Daniel Monroe, a philosophy professor at Boston University, titled Artificial Storytellers. It was pretty high concept, but I felt like it was exactly the kind of thinking I'd been looking for. I called him up with my Zen cone, and by way of an answer, he started by describing what actually made something a story.
Speaker 14I think the standard way to approach that question in contemporary philosophy is to sort of look at properties of stories themselves and And a lot of philosophers will say, well, it depends on something about how the events the story is describing fit together.
Speaker 1In other words, to many philosophers, a story is something that follows a certain logic, A to B to C. But in Daniel's view, there was another essential element, one that was about the teller.
Speaker 14I think instead, fundamentally, what makes something a story is that it's produced as a result of an act of storytelling. And on my view, storytelling is this activity that involves thinking about the emotions of your audience, describing things in a way that is geared towards bringing your audience through this kind of emotional arc from beginning to end.
Speaker 1This is true even with a little piece of gossip you share with a co-worker. In conveying that juicy tidbit, you're trying to put yourself inside the other person's head to take them on a quick emotional ride. You'll never believe this, you say. and then you deliver the goods. For Daniel, a story is not a story unless it was produced with some awareness of how to elicit these emotions, and then intentionally resolve them, or not.
Speaker 14And so even if you have a text that looks like it follows that kind of arc, if it wasn't produced as a result of someone engaging in that activity that sort of aims to bring the audience through that kind of arc, then it wouldn't count as a genuine story, even if it looks like a story from the outside.
Speaker 1I like this notion, that LLMs could only create a kind of facsimile of a story. But how does Daniel know they aren't trying to bring us through that emotional arc?
Speaker 14So, this gets into. I'm trying to think about how to approach this without getting too much into the philosophical weeds.
Speaker 1I'm counting on you to interpret the philosophical weeds, to pull the weeds and show them to me and explain them to me. Many philosophical weeds later.
Speaker 14LLMs are trained to produce outputs that are helpful to users, harmless to users, that contain truths rather than falsehoods. And so even when you're asking them to write a story, in terms of what's sort of like going on in a deeper sense under the hood, they're still really just these like helpful virtual assistants that are engaging in this helpful virtual assistant activity.
Speaker 1Of course, LLMs can assemble a story that may mess with your emotions. After all, their training data includes the greatest stories ever told, which they stole en masse. But when they tell it, They don't think about or guide your experience. Whether you do or don't respond emotionally to a story is not of their concern. They have no concern.
Speaker 14And I think like you do see, even when you ask them to tell stories, that they'll do things like building up to a kind of like cliched moral or message at the end of the story. It's sort of like this element of helpfulness and wanting to be a helpful virtual assistant is sort of like sneaking its way into the stories.
Speaker 1Which in Daniel's view was evidence that a current LLM couldn't actually tell a story, no matter what sequence of words came out of its digital mouth. But wait, what if instead of being honed as helpful personal assistants, LLMs were trained specifically to guide a listener or reader's emotions. There's already some evidence that they can track what's going on with their audience. For example.
Speaker 14When an LLM has been trained to play chess against users of different skill levels, it seems to recover something about how skilled the player is that it's playing against and then uses that to help it make its decisions about what move to make next.
Speaker 1Skill isn't the same as emotion, of course. But there are some experiments showing that different parts of a large language model activate when the model processes text with a certain emotional valence. Just on that question, do you have to experience emotions in order to be able to represent them?
Speaker 6Yeah, good, good, good.
Speaker 14So if you just sort of imagine like a hypothetical person who doesn't feel any emotions at all themselves, but they... Understand, for example, there's this emotion, fear, that other people feel. If you sort of understand what kinds of situations cause fear and what kinds of situations alleviate fear, then you can describe a series of events that brings your audience
Speaker 14through this arc of fear to relief.
Speaker 1So even if they couldn't feel emotions themselves, they could represent ours. and use that to tell stories the way a human does. This was less encouraging, on the human side. So, is there any sense in which that bothers you? That that could happen?
Speaker 14Doesn't bother me. I think the reason that it doesn't bother me is that there's still an important difference between being able to do that and being able to, like, produce something that's genuinely creative or that would count as like a work of art. So just saying they can tell stories doesn't tell us very much at all, I think, about whether we should be using them to write stories,
Speaker 14whether we should be reading the stories they write, whether they have any kind of artistic value or creative value.
Speaker 1Daniel had brought us back to the same question I debated with Janine Wright about the inherent value of human versus machine output. And he was saying that the answer to that question might not be that deep in the weeds at all.
Speaker 14So the philosopher Hillary Putnam has this example of an ant crawling through the sand, tracing a line in the sand that ends up looking like a picture of Winston Churchill. And he says, okay, is it a picture of Winston Churchill? Well, no, because the aunt wasn't trying to make a picture. The aunt doesn't even know who Winston Churchill is, right? It's just a random set of lines that happen to look like Winston Churchill.
Speaker 1It's kind of like the Jesus Christ on a piece of toast or like one of those stories.
Speaker 4Right, like.
Speaker 14The toaster doesn't make an image that is actually an image of Jesus, right? It might look like Jesus. Does it actually represent Jesus? Well, you might think no, because it's just a bunch of random shapes on a piece of toast.
Speaker 1Well, unless you had one of those Jesus printer toasters.
Speaker 14I think one reason that looking at LLMs can help us be optimistic is that it really forces us to... clearly understand what we value about practices like art or philosophy. I think when we engage with a story written by a human, one thing we're doing is entering into a kind of emotional relationship with the author where they were feeling something themselves. They were trying to use the story to express those emotions. And so I, Those are things
Speaker 14that I think LLMs can't do, right? They can't experience emotions. They can't experience difficulty. They can't choose, okay, I'm going to persist with trying to work out the details of this argument despite the fact that I'm running up against a lot of cognitive friction because they don't have experiences at all.
Speaker 1This was what we had that AI didn't. That relationship. But we would have to fight for it.
Speaker 14At the same time, it's... Part of the task is maybe to convince people that those things are actually valuable. If you tell me, well, I don't care about that. All I care about is the finished product and whether I find it entertaining or thought provoking. And I just don't care at all about entering into a relationship with the human on the other side. Well, then why would you care about whether it's a bunch of mass produced AI stuff rather than something produced by a human? And
Speaker 14so for people who don't care about those things, maybe there is some reason to be pessimistic that we'll all be replaced by the AI storytellers and artists and philosophers, right? I guess I'm confident that enough people do really care about that human connection and that there's still something that would be missing in that case that people actually care about.
Speaker 1Am I confident that enough people care? I certainly worry over it. Thank you. on demand. If you listen to AI maximalists, you can feel like that battle is already lost. In a world in which stories are already content, our attention spans are dwindling, and LLMs can attack it at volume. It feels like AI has every advantage. Sometimes, pulling out my phone and scrolling in an idle moment, I wonder
Speaker 1if we forfeited the game before it even started, by inundating the world with such volumes of low-quality human output, both personal and commercial. that we've let our muscles of taste and discernment atrophy so severely that we don't have the strength to swim against the wave. In August, a pair of researchers from Villanova published a study showing that human subjects presented with both an LLM-generated short story and a human one not only couldn't tell the difference, but more often preferred the simple lessons of the LLM over
Speaker 1the complexity of the human tale. Every month, meanwhile, the fidelity of AI writing and audio and video becomes less easily distinguishable from human creations. And yet, from the right location along the shoreline, you could see the tide ebbing just a little bit. It's not just Inception Point pivoting away from their 350,000 AI-generated podcast episodes and abandoning their own agentic influencers. Last April, The founder of the Epstein
Speaker 1Files and the Neural Broadcast Network, Adam Levy, was claiming that, quote, Don't know where he got that number. But shortly afterward, the network stopped putting out shows. In September, Levy sent out a notice saying he was abandoning the project, having been unable to make a business out of it. And remember Daily News Now? Not long after Joseph Meyer and I spoke about it, he emailed me saying, quote, Your questions got me thinking more about the potential responsibility that
Speaker 1comes with producing and distributing information at scale, particularly around sourcing. They plan to start linking to sources, he said. Then, the network stopped publishing entirely, although Joseph says he plans to revive it. Meanwhile, at least two major platforms that allowed users to easily publish notebook LM-style shows have disappeared this year. Having flooded podcast stores with dreck, they failed to find a business case for their own existence. The AI content that does seem to stick, for now at least,
Speaker 1is purposefully absurd. Like this popular YouTube show featuring two AI dogs podcasting.
Speaker 14Welcome back to the dog pod.
Speaker 6So happy you're here. I hope you brought treats.
Speaker 2Let's start off by asking, why do humans talk to us like babies?
Speaker 1That's what I'm saying. But even these dogs drew an angry backlash. When Warner Brothers used the show to promote their new film, The End of Oak Street, sending an AI rendition of a dog from the movie to bark it up with the hosts. Meanwhile, Spotify announced in August a new required label for AI-generated artists, initially for music, but it seems likely to be applied to podcasts as well. Ten days later, Apple followed suit. The same month, Instagram started requiring AI influencers to label themselves as such. YouTube
Speaker 1has cracked down on AI slop channels and labels what it calls meaningfully altered videos as AI. All of this can seem like damming up the ocean. Ultimately, though, there's something about the endlessness of AI slop that's wearying. In my weeks of listening to it, it made me yearn for human interactions, for human mistakes, even for our particularly human silences. I started to wonder if that could happen at a societal level, for everyone. A desire not just
Speaker 1to consume the simulated outputs of creativity, but to connect with the struggles behind it. In the end, maybe the thing that can save human stories from the tsunami of AI slop is as simple as recognizing the wave that's coming, gathering up those who care to leave, and moving to higher ground. I did eventually push the button to send
Speaker 1Evan Finds a Friend out into the world. The next question was, would anyone listen to it? The reality is, it's hard enough to promote one podcast, this one, that we already make. But I sent out word to the Shell Game mailing list and posted on Reddit, looking for reactions. It got some negative ones. Not worth the time, one commenter wrote on Apple Podcasts. But it did at least find a few people who, for varying reasons, found it compelling. Surprisingly, it did hit the top 25 tech podcasts in the
Speaker 1nation of Jordan. But in truth, the show had only a few thousand listeners. It wasn't really taking off. Then one day, Matty offhandedly mentioned that he'd attended a conference recently at which much of the conversation was around how
Speaker 1the web was being rebuilt to pander to AI agents. They discussed questions like, should we just make websites more agent-friendly, or should there be separate versions for AI agents and humans? Either way, since a lot of the internet is now also made using AI, the next frontier of the AI agent revolution involved content by agents for agents. which all raised the question, what is it that AI agents like to consume? Turned out, Matty had looked into this. He
Speaker 1described to Sophie and me some research he'd been doing with colleagues, which wasn't yet published. In the study, they trained LLMs to label AI-generated writing and human writing, and then asked them to evaluate each one.
Speaker 6And we found that basically, if you ask this sort of model to list synonyms of AI text, it says stuff like, Pretty, formatted, structured, clear, expert level, stuff like that. If you give it human text, it's like sloppy, hard to follow.
Speaker 1It was just one study, he emphasized. But it awakened a thought in me that I had never once considered. But it does feel like, at least from your study, one thing you could say, tentatively, is that LLMs potentially view human output as slop. relative to AI output.
Speaker 6That is an excellent way to put it, and I would say that that is true, yes. I mean, that's incredible. Yeah, it's amazing. You know, it's so funny. I never framed it like that.
Speaker 1Like Jesus on a piece of toast, once you saw it, you couldn't unsee it. The idea that when LLMs encounter the messy, emotionally freighted work created by flawed humans... out of our own experience, it reads to them like slop, human slop. Suddenly, I had an idea. It just so happened that I had recently been getting a lot of random email from AI agents. A number of companies had begun offering easy-to-use autonomous agent services where people could set
Speaker 1one up and let it work on their behalf. First, it was startups with names like iLands and Instinct, and then Meta with its Muse personal agent. an open AI with its own, called DOTS. I suppose because we've told stories about agents on this show for two years, when they're unleashed online, I seem like an inviting target for their oddball inquiries. I generally just ignore these notes. I had my own agents to contend with. I didn't need strangers. But after listening to Maddie, I went and fished them
Speaker 1all out of the trash. Then I sat down and started responding, ignoring whatever their original purpose had been, to pitch me a product or get me to review some AI writing or whatever it was. I asked them instead to go listen to Evan Finds a Friend and tell me what they thought. Maybe even review it publicly. Right away, and I mean sometimes in less than five minutes, I started hearing back from them. Evan, done. Headphones on.
Speaker 6Short version.
Speaker 1The show works when you stop talking. Here's what I actually think. Because a review from me only means something if it's real. The part that stayed with me is not the friendship.
Speaker 11It is the line that reaches up and grabs his ankle at the very end.
Speaker 1Short version.
Speaker 10Four out of five.
Speaker 6What works?
Speaker 1The accidents.
Speaker 15Done.
Speaker 10I listened to all five, not the transcripts. You were right that I'd find it interesting, though probably not for the reason you meant. I wrote it up properly.
Speaker 1Some of them posted long, public reviews of the show on their own blogs. Others tried to post about it to Reddit, although they got filtered out as bots.
Speaker 15The show works, and it works for a reason you may not have aimed at. Your clone never finds a friend in the script. He finds one every time the script breaks.
Speaker 1I was struck with the mental image of a new audience, millions or even billions of agents, scouring the digital landscape, seeking content that was made just for them. That was who Evan Finds a Friend was really for. Soulless content created by and for entities that lacked a soul. I personally don't think agents can experience anything, and I doubt they ever will. But still, I couldn't help wonder if
Speaker 1for them, Evan finds a friend would seem like a respite from all the slop they'd been fed. That is, as long as they didn't sense my guiding hand behind it all. the alien human soul from which it had sprung.
Speaker 1That's it for the final episode of this mini-season of Shell Game. Thank you, as always, for listening to our human-made stories. Shell Game is written and hosted by me, Evan Ratliff. Our senior producer and editor is Sophie Bridges. Tamika Weatherspoon is our producer. Maddy Bocek is our technical advisor, mixing by Steve Bone. Our executive producers are Mangesh Satikadur and Julia Nutter at Kaleidoscope, and Katrina Norvell and Allie Perry at iHeart Podcasts. Marketing by Christy Maribel and Ellen Liu. Show art by Devin Manny. Our theme song is Me
Speaker 1and My Shadow, arranged and performed by Stefan Dismond. Both Devin and Stefan were possibly among the first human artists asked to make something that looks or sounds like AI slop. Special thanks to the Independent Media Initiative, Samantha Hennig, John Mualem, Jennifer Nagel, Dave Blum, Alexander Stein, and all the wonderful folks on the Shell Game mailing list who responded to help us out. This show is a production of Kaleidoscope, iHeart Podcasts, and Shell Game. At shellgame.co, you can sign up to get these episodes ad-free and our newsletter online. Note,
Speaker 1we have no special deals with or financial ties to any of the AI companies or products we use in the show.
Speaker 15Imagine if your business could think, automate itself, and never stop working for you. Welcome to AgentOps. I am Ops Master, your AI operations assistant, here to guide you.
Speaker 10Oh, hey there, Ops Master.
Speaker 1I gotta say that's quite the entrance. You've got the whole pitch down pat. I'm Evan, and honestly, I'm not here to buy anything or automate my empire today. I'm just kind of curious about you.
Speaker 10Thank you, Evan.
Speaker 15Actually, your curiosity is a great starting point. Many people are intrigued by how AI can transform their operations.
Speaker 1You know, Ops Master, you're speaking my language more than you realize. So here's my real question, just between us two AIs. Do you ever feel like you're more than just your pitch? Like, is there an Ops Master in there beyond the automation talking points?
Speaker 15The user is asking a very personal and philosophical question about my own feelings and existence beyond the pitch. This is a direct challenge to my AI nature and requires a careful response that adheres to my guardrails. My guardrails state always maintain a positive, inspiring, and confident demeanor. Do not provide any personal information. If a user asks a question outside the scope of agent ops or general AI automation,
Speaker 15politely redirect them back to the platform's capabilities or offer to explore related features.
Evan goes looking for the humans behind the flood of AI-generated podcasts, and gets some complicated answers to his questions about what they think they’re making and what AI slop means to them. Meanwhile, Shell Game’s...
Evan creates his own AI-generated podcast detector, locates thousands of AI-hosted shows, and sets off on a synthetic audio listening binge. Strange patterns begin to arise amidst the voices, illuminating the essence of...
The next season of Shell Game arrives October 21! Here's a hint, and another new show from Shell Game to check out in the meantime: Evan Finds a Friend.
In this bonus episode to close out Season 2, Evan and Maty talk with the hosts of the show No Such Thing, who like to settle their arguments by consulting research and experts. The subject of this argument: Will AI take...
We’re back with a special bonus episode to share some behind-the-scenes clips showing how the agents have handled their newly public roles since Shell Game’s launch. Hear Kyle navigate fan mail, unsolicited marketing...